文献类型：期刊文献

英文题名：Assemble like expert: Multitask-meta hierarchical imitation learning algorithm guided by an expert skill model for robot polygonal peg-in-hole assembly

作者：Chu, Hubo[1];Zhang, Tie[2];Zou, Yanbiao[2];Sun, Hanlei[3]

机构：[1]Guangdong Ocean Univ, Sch Mech Engn, Zhanjiang 524088, Peoples R China;[2]South China Univ Technol, Sch Mech & Automot Engn, Guangzhou 510640, Peoples R China;[3]Shenzhen Res Inst Nankai Univ, Inst Intelligence Technol & Robot Syst, Shenzhen 518083, Peoples R China

年份：2025

卷号：83

起止页码：82

外文期刊名：JOURNAL OF MANUFACTURING SYSTEMS

收录：SCI-EXPANDED(收录号：WOS:001569388900001)、、EI(收录号：20253719133348)、Scopus(收录号：2-s2.0-105015304340)、WOS

基金：This work was supported by the National Major Science and Technology Project of China [grant number 2020YFC2007600] and the Natural Science Foundation of Guangdong Province [grant number 2024A1515012637] .

语种：英文

外文关键词：Expert assembly skill learning; Human behavior imitation; Robot automatic assembly; Meta imitation learning; Artificial intelligences

外文摘要：Robot polygonal peg-in-hole assembly is still challenging due to the unknown assembly environment and diverse tasks. To equip robots with expert assembly skills, this paper employs a model-guided strategy learning approach and proposes a multitask-meta hierarchical imitation learning algorithm guided by an expert skill model. Specifically, to construct a skill model for guiding strategy learning, a deterministic expert strategy is proposed. Based on this strategy, expert assembly characteristics are analyzed, and an expert skill model is developed to represent these characteristics. Furthermore, to learn experts' skill adjustment and generalization strategies across different tasks, a multitask-meta hierarchical imitation learning algorithm (MMHIL) is proposed. A parallel encoding attention network is designed to assist MMHIL in extracting multi-level skill information and learning assembly actions. A multitask-meta learning generalization framework with a mutual supervised learning optimization mechanism is proposed to enable MMHIL to rapidly adapt to new assembly tasks with limited training data. Comparative verification and polygonal peg-in-hole assembly experiments show that MMHIL has better skill learning effects and higher assembly success rates.